SD14 Exposure Analyzed

Steaphany

Well-Known Member
This was on my mind for a while. I did some exposure tests regarding how the SD14 performed when shooting IR photos to get a handle on just how sensitive the Foveon imager is in the IR region of the spectrum and how far off the SD14's meter was for such conditions.

Well, when I entered the measurements into a spread sheet, I ended up with some strange results which I had to wait for my replacement IR blocking filter before I could run some detailed exposure tests. I just completed my analysis of 45 exposure test imagess and I found some interesting results.

Since I can't post the actual OpenOffice spread sheet, here is a screen capture:

Exposure.jpg

First some background on how I gathered the data presented here.

I set a moderate light level in my office which would remain constant through out the duration of my shoot. I set up my X-Rite ColorChecker as the primary object in the field of view and used my 28mm DG EX Aspherical Macro lens. The SD14 was on a tripod to ensure that every frame had the same scene and the SD14 was set to Sun Light White Balance and ISO 100. The default exposure turned out to be 1/13th at an f/1.8. I then switched the exposure to manual setting 1/13th @ f/1.8 as my starting point.

I then shot a exposure for each shutter speed value from 1/13th through to 1/4000th. I then returned to 1/13th and proceeded to shoot from 1/13th through to 2.5 Seconds. Lastly, I shot another exposure at 1/13th.

This gave me a collection of 45 images, three being at 1/13th which served as a consistency check to ensure that the room lighting actually remained constant through out the shooting period.

I bulk extracted the JPG images from the X3F files so that they would be easy to work with and I could be sure that every image had exactly the same post processing treatment.

Using Photoshop, I opened each JPG and measured the luminance of the Neutral 5 gray square which ideally would yield 127 when photographed in ideal conditions. The measurements were made by selecting a region of the Neutral 5 Gray square, choosing Luminance, and reading the mean value measured.

I also verified these measurements in Paint Shop Pro Version 7.02 since this version has no color management and would only report luminance values from the actual file data.

Now, to the spread sheet.

Column A is the decimal shutter speed.

Column B is the Exposure Value departure from the SD14 metered 1/13th Second. Since the ISO and Aperture remained constant, the EV is solely based on the shutter speed. Note that I achieved a significant exposure range of -8.3 through to +5.0.

Column C is the Photoshop Mean luminance measurements of the Neutral 5 Gray Square.

For a point of reference, I adapted the Ansel Adams zone exposure system to the 0 to 255 scale, the center values of each of the 11 zones are listed in Column D.

The chart has the Exposure Value along the X axis and the Neutral 5 Gray square image value along the Y. The divisions of the Y axis correspond to the Image Values corresponding to the boundaries of the Ansel Adams zones.

The Violet line which spans EV values of -5 to +5 defines the Ansel Adams zones. Note that the Violet line crosses a vertical exposure value line at the center of each Image Value Zone.

In when using the Ansel Adams zone system in film photography, an ideal exposure would span -3EV to +3EV for the bulk of the image with high lights and deep shadows extending into the -4EV and +4EV ranges. Exposures extending into the -5EV and +5EV are areas of the image which have no recognizable detail.

Note that I have 11 shades of Gray to the right of the chart whose values correspond to the zone center image value.

The SD14 exposure performance is shown by the Blue line.

The first thing to note is where the SD14's exposure line crosses the 0EV point. The SD14's metered exposure is slightly shifted to the right, instead of 127, the Neutral 5 gray square yield a image value of 138.

Second, the slope is actually more drastic than the Ansel Adams zone line meaning that the SD14 would yield a more contrasty image.

Third, the SD14's exposure is not symmetrical about the 0EV point. Moving towards a brighter image value, the SD14 is nicely linear but it will saturate at only +3EV. If you ever use a separate spot exposure meter, any object brighter than +3EV from the SD14's settings will be washed out. Moving from the 0EV point towards the darker image values, the SD14's exposure is only linear to about -2EV. The Red straight line is a linear regression line of the measured image values over -2EV through to +3EV showing that this region of the SD14's exposure curve is truly linear. Beyond -2EV, the SD14's exposure plateaus. Where the linear regression predicts the darkest image value zone being achieved at -3EV, in reality that point is not reached until going as far as -8EV.

This dark image value plateau also explains why the SD14 is prone to noise in low light or in dark shadows.

How to work with this information.

First off, I'm going to set my exposure compensation to -0.3EV to bring objects with Neutral 5 gray values closer to the desired 127.

When shooting to the right, It's best to be conservative on how far you go. With only 3EV of head room, it would not take much of a bright scene to end up with loss of detail in the high lights.

Watch the darks and shadows. Since the SD14 exposure is not symmetric between the positive and negative values from 0EV, it really would not take much to become "too dark".
 

foveonfan

Well-Known Member
Whew, Steaphany!

You ARE a busy little beaver! Quite some work there. It's amazing what methodical people can achieve. As for this "pinball" machine between my ears, it will take me some time to reasonably digest most of this valuable information. Thanks for sharing.

Sincere regards, Jim Roelofs

Please visit my humble gallery @
Please, Log in or Register to view URLs content!
 

Steaphany

Well-Known Member
Hmmm, I didn't feel this was too techi...

Besides, Pinball machines are fairly sophisticated, an interacting mechanical system biased by subtle manifestations of chaos, and we all know how the structure of everything has it's roots in chaos.

Well, here is a better visualization that I'm sure will have everyone scratching their heads with a loud group chorus of "Huh !!":

Histogram_Explaineds.jpg

Yes, the famous Sigma SD14 histogram, now with numbers measured in EV. To the right of Center all looks fairly calm going from 0EV, +1.5EV, +3EV for the scale of image values 127, 197, and 254. High strangeness abounds left of 0EV with scale going from 0EV, -2EV, -8EV for image values 127, 58, and 0.

How does this impact your photography ? Well, for example, your subject of a photo corresponds to a peak at the Left Tick. (The Red and Green channels have peaks here in the image above) You want the exposure to bring your subject to mid range. Now that you know that the Left Tick is at -2EV, an exposure compensation of +2EV will get your image right where you want it. But, be warned, if the right most data in the histogram extends much beyond the Right Tick, (Also the situation shown in the above image) which corresponds to +1.5EV, a +2EV exposure compensation will shift values at the Right Tick out to +3.5EV. All the image detail that gets shifted beyond the edge of the SD14 world of +3EV will be washed out.

Another example would have a photo subject corresponding to a peak at the Right Tick. This is easy, a -1.5EV exposure compensation would bring the subject precisely to mid range. If shadows and darks were left of the Left Tick, exposure wise you'll be Ok since you can get to -8EV, but be watchful of image noise.
 

Arvo

Well-Known Member
Very impressive work!

Results are somewhat interesting (or weird).
Usually digital cameras response curve is linear in central part and flattens at both ends (S-like).
Lower part of foveon curve corresponds to sensors usual behavior (maybe tail is longer), but upper part not.

This strange response curve may be one cause of so-called "Foveon look".

Interesting, how would these curves look in pure raw data? (I understand that you used camera jpgs?) If you had some good gradient image, it would be sufficient to take one-two overexposed shots from it and look into raw data, presumably channel-wise :)

From other side your test shows that Foveon sensor dynamic range is up to about 11 stops - not bad.
 

foveonfan

Well-Known Member
Arvo, the dynamic range of the Foveon sensor was one of the main reasons I chose it as my preferred digital photo technology.

Steaphany, if you were disappointed in my reference to the highly technical nature of your original post, it was definitely not intended as such. It was meant as a self deprecation, along with the "pinball" reference. But your high regard for this instrument gives me hope. ;)

My personal intelligence has suffered somewhat form the after effects of a couple of medical conditions that tried to lay me low. I've even had to reach down into the bowels of my "pinball machine" to retrieve some of the knowledge and re-learn some old skills. But - and isn't there always a but? - with your timely encouragement and a continuing will to recover, who knows where the limit of the sky will be this time around. :)

Yes, the dreaded histogram used to scare me wittless but these days I have made it my friend too.

I really appreciate your sharing of your experiences in this fashion, Steaphany, I find it enlightening and helping to keep me "on track" with learning something every day.

Sincere regards, JR

Please visit my humble gallery @
Please, Log in or Register to view URLs content!
 

Steaphany

Well-Known Member
Jim,

I understood your pinball machine reference. I'm also glad to hear that you appreciated my comments. It's also good to hear that your motivated, and my techi post encourage, your continued recovery from medical issues.

As you can see from my follow up post, I am aware that not everyone can look at a chart or formula and understand how to put the information into use.
 

Steaphany

Well-Known Member
RAW Exposure for Arvo

Since Arvo asked, here it is, the SD14 exposure curve taken from RAW data.

First, my method.

I extracted 48 bit per pixel Tiff format files from the original X3F files using X3F.exe with the transformation matrix:

Code:
Please, Log in or Register to view codes content!
Each file was opened in Photoshop where I had a saved selection which I loaded into the open images files allowing the same number of pixels to be sampled at the same location within the image. This was necessary since many images were so dark, no details could be seen with out imposing some form of adjustments. I never did any adjustments to any of the tiff files, I simply loaded the selection and recorded the mean values for both the RGB and Luminosity histogram channels.

Here is a Photoshop screen capture as an example:

photoshop.jpg

and this is the data and chart from my open office spread sheet:

RAWSpreadSheet.jpg

Some items to note:

The dynamic range, especially for the RGB channel is far greater in RAW than unaltered JPG, exceeding 13EV.

The Scale of values along the Y axis are the values taken directly from Photoshop and I am not familiar with translating these to a known unit for comparison.

It is interesting to see that the Foveon imager is somewhat non-linear and I find it strange that there is a discontinuity, a kink, at the +4EV point of the curve.

To Arvo,

If you want my source X3F files for your own analysis and coding efforts, just let me know.
 

foveonfan

Well-Known Member
One question. How can there be an RGB channel when R,G and B are separate colour channels?

Sincere regards, JR
 

Steaphany

Well-Known Member
If your Photoshop does not show these wonderful measurements and calculations, there is a small triangle on a small tab to the upper right of the histogram panel. It marks a pull down menu where you can select the extended Histogram and also select whether you want to see the statistics. I find these to be very useful to have on all the time.

Now to the Extended Histogram with statistics:

Photoshop has several sources of information when calculating the extended data from a histogram.

When there are no selections, the histogram and the statistical calculations cover the whole image. When a selection exists, only the selected area is used to determine the statistical calculations.

Along with whole image or a selection, there are several options listed as Channel from which the statistics are calculated. These being:

  • RGB Combined
  • Red, Green, and Blue individually
  • Luminosity
  • Colors
.

From these, Photoshop will compute:

  • The Mean pixel value
  • The Standard Deviation (Useful to measure noise - a pure region of an image will have a low standard deviation when the noise is low and a higher standard deviation when the same area is noisy)
  • The Median pixel value
  • The number of pixels used in the statistical calculations.

To actually get numbers from the histogram curve itself, position your cursor on the curve and you'll see that you can also get:

  • The Level (The cursor marking the point along the X Axis)
  • The Count (The height of the histogram at the cursor position)
  • The Percentile (What proportion of the image pixels have this value)
  • The Cache level (I'm not sure how this is used)

Now left click on the histogram and while holding the mouse button down, drag along a region of the histogram. Now the Level will be the range selected, the Count will be the sum of the histogram curve for the Level range, and Percentile will be the percentage of the image pixel values selected. This histogram selection will vanish as soon as you release the mouse button.

So, expand your histogram, turn on the statistics, and have fun getting numbers from that curve.

FYI, I am not a Photoshop guru. In fact, I've only got Photoshop after getting my SD14, so I still have a lot to learn. I found these extended and statistical features of the Photoshop histogram when I needed a way to numerically assess SD14 images for noise to permit me to better understand how well various techniques caused and reduced noise. I did not want a visual critique or feeling whether the noise went up or down, but actual numbers which measure levels of noise and just how much of an improvement could be gained.
 

sendhil76

Active Member
Late reply to this post...BUT Steaphany...awesome work!!

This is the best information i have read till date regarding the Foveon imaging sensor and its share of exposure quirks.

Thanks very much for taking the time and effort to figure this all out and sharing it here! Very much appreciated!

Sendhil.

P.s: After taking lots of pictures and going the trial and error way...i figured a +0.3 EV compensation gives me images with all the detail i need :) ...the other thing i need to figure is how to reduce noise in the image captures in low light...particularly when ambient temperatures are ~35 degrees celcius or more...
 

tc95

Well-Known Member
Wow....Steaphany...Great stuff....I have been away at work...and came back and read your post....I can't wait to get my Camera back from Sigma...they are checking to see if they can calibrate-it...I am having issues with the AF and MF focus on the camera.....I have done all the tweaks I can....

Thank you for sharing your info...

Tony C.
 

Arvo

Well-Known Member
Sorry for answering late - I somehow missed your original posting :(

It is interesting to see that the Foveon imager is somewhat non-linear and I find it strange that there is a discontinuity, a kink, at the +4EV point of the curve.
It may be some colorspace conversion effect. Lately, dealing with one red flower, I had problem with out-of-gamut colors and clipping these could create similar effects. RGB part of graph appears to go into saturation at approximately same point too.

About nonlinearity - did PhotoShop know that your tiff is *linear*? If not, then it probably applied (reverse) gamma correction to image.

13EV dynamic range is near impossible; sensor official range is a bit over 10EV and your JPG tests seem confirm exactly that number.

To Arvo,

If you want my source X3F files for your own analysis and coding efforts, just let me know.
Yes, I want same source files of course :)
 

Arvo

Well-Known Member
I played a bit with Steaphany files. Graph below is purely technical (like my notes), but maybe it does help to someone, wanting better understand Foveon sensor characteristics. At least usable EV range can be seen on that graph easily.

I analyzed pure RAW data before conversion it to real RGB values; that means that RGB values on graph are sensor readings and they do not exactly correspond to visible colors.

On graph there are two families of response curves - for colorchechker card white (upper curves) and darkest gray (lower curves) patch. EV values are relative to some 'normal' exposure; I think this 'normal' is actually a bit underexposed, but this doesn't change sensor behavior.

I added error bars, displaying standard deviation of measured values; they may be used as 'noise' indicator. (Note that 'noise' wouldn't be computable if I would use entire image for data source.) Using symmetrical deviation is actually not correct, for lower values error margins went below zero, what is impossible - thereby take that 'noise' as only very rough approximation of real noise. I din't analyze 'noise' characteristics.

Because human eye sensibility is logarithmic, then RAW values are presented on logarithmic scale. EV values are internally logarithmic too - one stop means 'two times'.
I think such logarithmic behavior of RAW values was not taken into consideration in previous graphs, thereby they are not so linear :)

RAW internal values usually max out about at 6600, although for gray patch blue channel got as high of 9000. Most probably sensor data is altered before writing into RAW file; I've somewhere read about linearization and dark frame is subtracted too. A/D converter in camera rumoured to be 12bit one; looking at graps there can be 12bit (=12EV) of dynamic range (from -8 to +4EV), but noise makes lower part of it unusable. 10 stops seems achievable; depends on noise characteristic.

log-raw-ev.jpg

If someone is interested in actual numerical values (and excel table), please email me or ask here.
 

Steaphany

Well-Known Member
I find your peak values interesting since the SD14's 12 bit ADC should only yield values between 0 and 4095 inclusive.

Obviously, to get anything above 4095 means that some form of scaling is going on.
 

sendhil76

Active Member
Nice work Arvo! And i will be interested in having a look at your Excel file!

Please mail it to me at: sendhil76@gmail.com

And the image pipeline processor used by Sigma SD14 is the Analog Devices Blackfin ADSP-BF561 series.

Search for "Sigma SD14 Analog Blackfin" in google and you will get relevant links on the first page itself.

If interested in the processor datasheets, please let me know and i will lokk them up in my archive and mail them to you!

Thanks and stay well!
Sendhil.
 

Steaphany

Well-Known Member
Everything your could ever want to know about the Analog Devices ADSP-BF561 is here:

Please, Log in or Register to view URLs content!


Working in the field of Electronic Engineering for the past 33 years, I already have most semi companies book marked. :)
 

Steaphany

Well-Known Member
The ADSP-BF561 is a nice processor, but it provides no on chip ADC functionality. It does provide two parallel data channels to interface to external ADC video data sources, but that gives no clue on the actual ADC's that Sigma interfaced with the Blackfin.

This is still good to know to help my efforts to decompile the SD14 1.08 software installables.
 

sendhil76

Active Member
Pleased to know about you Stephany :) ...i am hovering around 1/3 of your experience :) LOL

Yes, totally agree with you...what i hold in my hand to shoot pictures with is really a 'black box' ...and a very dark one too :) ...the suspense is killing me...and yet that is exactly what keeps experimenting with it all the time!

It would be nice if i can get my hands on a 'brick' SD14 to have a look inside at the engineering and design...a lot of the questions i have inside my head will get answered...and my soul will get some badly needed rest :)

Thanks for your posts!
 
Top